Running multiple NVIDIA GPUs for AI workloads in Proxmox VE can cause early boot hangs if the host OS tries to load conflicting drivers. In this guide I document how my Proxmox host with 2× RTX 3090 was stuck at systemd-modules-load, how I debugged it, which files to inspect (/etc/default/grub, /etc/modprobe.d/, /etc/modules-load.d/), and the final stable configuration for rock-solid GPU passthrough to an Ubuntu VM.

I built on my edge AI hardware by adding Proxmox VE as the virtualization layer. After prepping BIOS, using Rufus with the nomodeset trick, and installing Proxmox, I enabled IOMMU, configured VFIO, and passed through 2× RTX 3090 GPUs to a single Ubuntu VM. This setup lets me run private AI workloads at near bare-metal speed, while keeping Windows and native Ubuntu for special use cases.

Exploring three budget AI supercomputer paths: a Dell R740xd for enterprise labs with big storage but limited GPU flexibility, a TRX50 + Threadripper 7970X workstation offering fast DDR5, Gen5 NVMe, and dual RTX GPU power, and the futuristic GB10 AI desktop with unified CPU/GPU memory. Dell is lab-friendly, GB10 is AI-only, but the TRX50 build strikes the best balance today.